· via dev.to (home feed)
Google ships Gemini 4 Neon, a frontier model at $3 per million input tokens
Google's fourth Gemini model in seven weeks runs at roughly twice Carbon's speed for $3 per million input tokens, but an independent test flags steep latency and cost gaps for high-volume automation.

Google released Gemini 4 Neon on November 18, a frontier-tier model aimed at completing tasks end to end quickly in software engineering, knowledge work and science. According to the announcement from Koray Kavukcuoglu, SVP at Google DeepMind, the new model sits between Carbon, designed for heavyweight long-running projects, and Hydrogen, designed for fast day-to-day execution.
The release pace itself is notable. As tracked by Latent Space and Arena, the gap between Gemini releases has shrunk from 28 days (Argon to Carbon) to 15 and then 6, making Neon the fourth model in the family within roughly seven weeks. Argon has been retired.
Specs and pricing
The AI Studio documentation lists the model ID as gemini-4-neon. Inputs are text, image, video, audio and PDF; output is text only. The context window is 1,048,576 tokens, default maximum output is 65,536 tokens, and a long-decode option extends output up to 1M tokens.
Pricing is $3 per million input tokens, $15 per million output tokens and $0.15 per million cached tokens, a 95% cache discount. One wrinkle: Carbon is on a promotional rate of $2 and $10 per million tokens until December 1, after which it moves to $4 and $20. Measured against that promotion, Neon is 50% more expensive.
Benchmarks and rankings
Google reports a score of 61 on the Artificial Analysis Intelligence Index at xhigh, good for third place overall; second on the Vals Index at 70.9%; a second-place Agent Arena debut at 1336; first on the Coding Agent Index at 58.6; and first on Zapier AutomationBench at 56.3%. On coding, Google cites 12.8% on SWE-sweep using the mini-SWE harness and a state-of-the-art 39.0% inside Google Antigravity, along with 99.2% agent honesty and a vendor-claimed 100% detection of indirect prompt injection on Gray Swan IPI.
The Decoder cautions that third is not first: GPT-6.2 Astra and Claude Fable 5.2 rank above Neon on the index and are absent from Google's comparison table.
Speed and per-task cost
According to Artificial Analysis data cited by The Decoder, Neon generates around 110 tokens per second versus 57 for Carbon, roughly twice as fast, cutting a benchmark run from 36 minutes to about 21. The savings, however, come from list price rather than token efficiency: Neon uses about 15% more output tokens per task, landing at $6.44 against Carbon's $7.76.
For a coding task in Antigravity, The Decoder and Artificial Analysis put Neon at $6.86 per task, compared with $12.61 for Claude Opus 5.5 and $15.47 for Claude Fable 5.2.
Availability and integrations
Neon went live on November 18 in the Gemini API paid tier within Google AI Studio and Android Studio, in Gemini for business, in Antigravity for AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra users. Latent Space reports day-one integrations with Cursor, GitHub Copilot and LangChain, and Simon Willison notes that the llm-gemini 0.38 release added Neon support to the llm CLI on the same day.
Per Google, confirmed by InfoWorld, Neon also operates inside Antigravity as an Orchestrator that directs parallel Hydrogen subagents on multi-file debugging and refactoring.
A caution for automation workloads
An independent test by Activepieces complicates the picture for automation pipelines. Neon completed 3 of 3 tasks against 2 of 3 for GPT-5.4 mini, but median latency was 6.4 seconds versus 1.0, and per-task cost came out roughly 54 times higher, near $8.30 versus $0.15 at 1,000 tasks per day. Activepieces' conclusion is to validate price and latency on your own tasks before moving latency-sensitive or high-volume workflows.
Why it matters
Two things stand out. First, the cadence: four frontier-tier models in seven weeks, with older ones retired, signals that Google now treats model releases as a rolling pipeline rather than annual events, and that undercutting rivals on list price is part of the playbook. Second, the gap between headline and effective cost: $3 per million input tokens looks cheap, but per-task economics depend on token efficiency, latency and volume, and the Activepieces result shows a frontier planner can cost dozens of times more than a small model on simple automated steps. The Orchestrator pattern, with an expensive model planning while cheap subagents execute, is emerging as the practical middle ground, and it appears to be where Google is steering the lineup.
- #gemini
- #llm
- #api-pricing
- #ai-agents